Cloudflare switches to EPYC 9684X Genoa-X CPUs with 3D V-Cache — 145% faster than previous-gen Milan servers
3D-VCache is Cloudflare's best friend
Cloud provider Cloudflare has upgraded its server infrastructure with AMD’s latest generation EPYC Genoa-X processors armed with 3D V-Cache technology. Cloudflare reports a 145% improvement in performance and a 63% improvement in efficiency over its previous generation servers, which were also EPYC-powered.
The new "Gen 12" servers come with AMD’s EPYC 96-Core 9684X Genoa-X processor featuring over one gigabyte of L3 cache. The system features 384GB of DDR5 4800MHz DDR5 memory operating in a 12-channel memory configuration, dual E1.S NVMe Samsung PM9A3 7.68TB or Micron 7450 Pro 7.68 TB SSDs, dual 25GbE Intel or Nvidia Mellanox networking, and a 800W titanium grade PSU.
Before focusing on the 9684X, Cloudflare conducted an extensive survey of the CPU landscape to determine which model would best suit its needs. The cloud service provider purportedly evaluated “many candidates in the lab” but ultimately shortlisted three AMD CPUs for evaluation, knocking out any Intel Xeon CPUs.
The three in question were the EPYC 9654, EPYC 9754, and the EPYC 9684X, which the provider eventually selected. The EPYC 9654 is the vanilla (non-X3D) variant of the 9684X, sporting 96 cores, 192 threads, 384MB of L3 cache, and a max turbo clock of 3.7GHz. The EPYC 9754 is a Burgamo Zen 4c powered CPU featuring 128 cores, 256 threads, 256MB of L3 cache, and a max boost clock of 3.1GHz. The 9684X features the same specs as the 9654 but with 1,152GB of total L3 cache.
These chips were evaluated against Cloudflare’s previous 11th-gen servers powered by the 64-core AMD EPYC 7713. The 9654, 9754, and 9684X were tuned to their max 400W configurable TDPs; the EPYC 7713 maximum TDP is 240W. All the chips were effectively benchmarked under a real-world workload, with each server set up to run the same processes and services as its production servers.
Cloudflare discovered that all three chips were at least 2x faster than the previous EPYC 7713 CPUs. The Genoa 9654 was exactly 2x faster than the 7713, the Bergamo 9754 was 2.15x quicker, and the Genoa-X 9684X was 2.45x faster. In terms of efficiency, Cloudflare discovered that the Genoa 9654 was 33% more power efficient than the EPYC 7713, the Bergamo 9754 was 38% more efficient, and the 9684X was the most efficient of the bunch, being 63% more power efficient.
Cloudflare provided specific details on the performance enhancements offered by AMD's 3D V-Cache technology, noting a massive 22.5% performance delta between the 9654 and 9684X in the tests it ran. The service provider said its workload benefits from more accessible low-level caches, avoiding trips to DDR5 memory, which introduces significant system latency.
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
The 9684X also delivered this 22.5% performance improvement within the same power envelope as the 9654. The 9684X's 3D-stacked L3 cache purportedly consumes additional power, but the difference only affected maximum all-core turbo clock speeds by just 3%.
With these results, it's obvious why Cloudflare opted for the 96-core EPYC 9684X for its 12th-gen servers. Its particular applications take full advantage of the extra cache offered on AMD's 3D V-Cache-equipped Genoa-X CPUs, boosting performance more than even the EPYC 9754 with its 128 cores.
Aaron Klotz is a contributing writer for Tom’s Hardware, covering news related to computer hardware such as CPUs, and graphics cards.
-
Udyr The system features 384GB of DDR5 4800MHz DDR5 memory operating in a 12-channel memory configuration
The system features 384GB of DDR5 4800MHz memory, operating in a 12-channel configuration.
and a 800W titanium grade PSU.
and an 800W titanium grade PSU. -
DS426 Wow, is it really worth CloudFlare switching out gen-on-gen servers? AMD will be announcing Zen 5-based 'Turin' EPYC before long.Reply
Nah. I think they're just pointing out that these will be their new server CPU's going forward, now with a known quantity on gains for comparison.
3D V-Cache does use more energy but less than additional trips to and from DDR5, which this case illustrates when that cache can get heavily utilized. -
Urzu1000 DS426 said:Wow, is it really worth CloudFlare switching out gen-on-gen servers? AMD will be announcing Zen 5-based 'Turin' EPYC before long.
Nah. I think they're just pointing out that these will be their new server CPU's going forward, now with a known quantity on gains for comparison.
3D V-Cache does use more energy but less than additional trips to and from DDR5, which this case illustrates when that cache can get heavily utilized.
You can see their value from the upgrade by the amount of RAM they're installing on these. 384GB of RAM is not a lot for servers of this size - something with a top end Genoa X CPU would typically ship with terrabytes of RAM.
This indicates that their workloads are very CPU intensive, and that they would likely benefit considerably from the enhanced performance and cache density.
Being an infrastructure company, their ongoing running costs are most likely astronomical and getting 145% more performance for similar power envelope means that they can cut their operating costs in half by replacing all their current servers with half of that number. On the flipside, it may also double their capacity for the same operating cost if they keep the same number, allowing them to generate more revenue for the same operating cost.
That's a huge incentive for them to quickly but these and liquidate their old servers by selling them, or repurposing them elsewhere within the company. -
Kondamin i doubt they will replace the old ones before they are broken or a write off.Reply
Players like these buy thousands of servers yearly just swapping recent machines out because of a slight performance improvement isn’t worth the man hours. -
Vanderlindemedia Kondamin said:i doubt they will replace the old ones before they are broken or a write off.
Players like these buy thousands of servers yearly just swapping recent machines out because of a slight performance improvement isn’t worth the man hours.
If the swap gains 145% then yes you need less overall for the same performance, less power and thus more efficient. I swapped out a bunch of XEON's in favor of Epyc. I could take down 3 servers with that alone.
williamcll said:Wondering who they are going to sell the old servers too?
they have their channels. High end hardware still sells. Cloudflare (i use it) is worth the money if you need it for web. -
Kondamin Vanderlindemedia said:If the swap gains 145% then yes you need less overall for the same performance, less power and thus more efficient. I swapped out a bunch of XEON's in favor of Epyc. I could take down 3 servers with that alone.
they have their channels. High end hardware still sells. Cloudflare (i use it) is worth the money if you need it for web.
its very likely they have a bunch of older servers running that require replacement before they start swapping machines that aren't 2 years old.
They have 330 data centres world wide, they need tens of thousands of these gen 12 servers, servers that need to be built by someone which takes time.
It is an interesting question. -
Vanderlindemedia Im sure the stuff is streamlined as much as possible. CF has been around for a long time and is expert at DDOS and much more other security based features.Reply
its a win for AMD as well knowing their whole park of DC's is swapped out with EPYC's.